Evaluation of Deep Learning and Conventional Approaches for Image Steganalysis

Abstract

Steganography is the technique for embedding secret messages into digital media without changing their appearances. As a countermeasure to steganography, steganalysis detects the presence of hidden data in digital content. For the last decade, the majority of image steganalysis approaches can be formed by two stages. The first stage is to extract effective features from the image content and the second is to train a classifier in machine learning by using the features from stage one. Ultimately the image steganalysis becomes a binary classification problem. Since Deep Learning related architecture unifies these two stages and saves researchers lots of time designing hand-crafted features, the design of a CNN-based steganalyzer has therefore received increasing attention over the past few years. In this paper, we will examine the development in image steganalysis, both in the spatial domain and in the JPEG domain, and discuss the future directions.